What We Mean When We Say AI-Assisted, Not AI-Led

Language matters in healthcare. When we say AI-assisted, we are not using a marketing qualifier. We are describing an architectural reality: the AI prepares, and the physician decides. Every clinical output that reaches a patient or a clinical record has passed through physician review. This is not aspirational — it is how the system is built, at the code level.

The alternative — AI-led care — is a system in which the AI's output is the care. A recommendation becomes a treatment plan without physician review. A flag becomes an intervention without clinical judgment. This model is faster. It is also dangerous, and we believe it is wrong. Not because AI is unreliable in some general sense, but because metabolic medicine involves irreducibly complex individual judgment that AI cannot replicate.

The physician brings something the AI cannot. A patient who reports fatigue might be reporting the expected side effect of a new medication, or the early presentation of something unrelated, or a response to a stressor in their personal life that is not in any clinical record. The physician who knows the patient can distinguish these. The AI cannot — not reliably, not safely, not yet, and perhaps not ever for some of these distinctions.

What the AI can do is reduce the preparation burden so the physician has more time for judgment. That is the correct framing. Not AI reduces the physician's workload — which implies substitution. AI increases the quality and reach of physician judgment by taking on the work that doesn't require it: data assembly, population screening, encounter preparation.

We are explicit about this because the distinction matters for what we're building. An AI-led system optimizes for throughput. An AI-assisted system optimizes for the quality of physician attention. Those are different products with different architectures and different outcomes. We chose the latter, and we designed the system so it cannot drift toward the former without a fundamental rebuild.